From 3fc3c4c4345b0a86508ab78103c3cae36b8de285 Mon Sep 17 00:00:00 2001 From: Jael Gu Date: Fri, 28 Jul 2023 14:32:19 +0800 Subject: [PATCH] Add files Signed-off-by: Jael Gu --- README.md | 140 ++++++++++++++++++++++++++++++++++++++++++++++- __init__.py | 5 ++ llama2.py | 90 ++++++++++++++++++++++++++++++ requirements.txt | 2 + 4 files changed, 236 insertions(+), 1 deletion(-) create mode 100644 __init__.py create mode 100644 llama2.py create mode 100644 requirements.txt diff --git a/README.md b/README.md index 3dba710..da6af99 100644 --- a/README.md +++ b/README.md @@ -1,2 +1,140 @@ -# Llama-2 +# Llama-2 Chat + +*author: Jael* + +
+ +## Description + +A LLM operator generates answer given prompt in messages using a large language model or service. +This operator uses a pretrained [Llama-2](https://ai.meta.com/llama) to generate response. +By default, it will download the model file from [HuggingFace](https://huggingface.co/TheBloke) +and then run the model with [Llama-cpp](https://github.com/ggerganov/llama.cpp). + +This operator will automatically install and run model with llama-cpp. +If the automatic installation fails in your environment, please refer to [llama-cpp-python](https://github.com/abetlen/llama-cpp-python) for instructions of manual installation. + +
+ +## Code Example + +Use the default model to continue the conversation from given messages. + +*Write a pipeline with explicit inputs/outputs name specifications:* + +```python +from towhee import pipe, ops + +p = ( + pipe.input('question', 'docs', 'history') + .map(('question', 'docs', 'history'), 'prompt', ops.prompt.question_answer()) + .map('prompt', 'answer', ops.LLM.Llama_2('llama-2-7b-chat')) + .output('answer') +) + +history=[('Who won the world series in 2020?', 'The Los Angeles Dodgers won the World Series in 2020.')] +question = 'Where was it played?' +answer = p(question, [], history).get()[0] +``` + +*Write a [retrieval-augmented generation pipeline](https://towhee.io/tasks/detail/pipeline/retrieval-augmented-generation) with explicit inputs/outputs name specifications:* + +```python +from towhee import pipe, ops + + +temp = '''Use the following pieces of context to answer the question at the end. + +{context} + +Question: {question} +''' + +system_msg = 'Your name is TowheeChat.' + +q1 = 'Who are you?' +q2 = 'What is Towhee?' + +p = ( + pipe.input('question', 'docs', 'history') + .map(('question', 'docs', 'history'), + 'prompt', + ops.prompt.template(temp, ['question', 'context'], sys_message)) + .map('prompt', 'answer', + ops.LLM.Llama_2(max_tokens=200)) + .output('answer') +) + +history = [] +docs = [] +ans1 = p(q1, docs, history).get()[0] +print(q1, ans1) + +history.append((q1, ans1)) +docs.append('Towhee is a cutting-edge framework designed to streamline the processing of unstructured data through the use of Large Language Model (LLM) based pipeline orchestration.') +ans2 = p(q2, docs, history) + +print(q2, ans2) +``` + +
+ +## Factory Constructor + +Create the operator via the following factory method: + +***LLM.Llama_2(model_name_or_file: str)*** + +**Parameters:** + +***model_name_or_file***: *str* + +The model name or path to the model file in string, defaults to 'llama-2-7b-chat'. +If model name is in `supported_model_names`, it will download corresponding model file from HuggingFace models. +You can also use the local path of a model file, which can be ran by llama-cpp-python. + +***\*\*kwargs*** + +Other model parameters such as temperature, max_tokens. + +
+ +## Interface + +The operator takes a piece of text in string as input. +It returns answer in json. + +***\_\_call\_\_(txt)*** + +**Parameters:** + +***messages***: *list* + +​ A list of messages to set up chat. +Must be a list of dictionaries with key value from "system", "question", "answer". For example, [{"question": "a past question?", "answer": "a past answer."}, {"question": "current question?"}] + +**Returns**: + +*answer: str* + +​ The answer generated. + +
+ +***supported_model_names()*** + +**Returns**: + +A dictionary of supported models with model name as key and huggingface hub id & model filename as value. + + { + 'llama-2-7b-chat': { + 'hf_id': 'TheBloke/Llama-2-7B-GGML', + 'filename': 'llama-2-7b.ggmlv3.q4_0.bin' + }, + 'llama-2-13-b-chat': { + 'hf_id': 'TheBloke/Llama-2-13B-GGML', + 'filename': 'llama-2-13b-chat.ggmlv3.q4_0.bin' + } + } diff --git a/__init__.py b/__init__.py new file mode 100644 index 0000000..38e9485 --- /dev/null +++ b/__init__.py @@ -0,0 +1,5 @@ +from .llama2 import LlamaCpp + + +def llama_2(*args, **kwargs): + return LlamaCpp(*args, **kwargs) diff --git a/llama2.py b/llama2.py new file mode 100644 index 0000000..86db3a0 --- /dev/null +++ b/llama2.py @@ -0,0 +1,90 @@ +# Copyright 2021 Zilliz. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +from typing import List + +from huggingface_hub import hf_hub_download +from llama_cpp import Llama + +from towhee.operator.base import PyOperator, SharedType + + +class LlamaCpp(PyOperator): + '''Wrapper of Dolly inference''' + def __init__(self, + model_name_or_file: str = 'llama-2-7b-chat', + **kwargs + ): + self.kwargs = kwargs + supported_model_names = self.supported_model_names() + + if model_name_or_file in supported_model_names: + model_info = supported_model_names[model_name_or_file] + hf_id = model_info['hf_id'] + model_filename = model_info['filename'] + self.model_path = hf_hub_download(repo_id=hf_id, filename=model_filename) + else: + self.model_path = model_name_or_file + assert os.path.isfile(self.model_path), f'Invalid model path: {self.model_path}' + + print(111, self.model_path) + self.model = Llama(model_path=self.model_path) + + def __call__(self, messages: List[dict]): + prompt = self.parse_inputs(messages) + resp = self.model(prompt, **self.kwargs) + answer = self.parse_outputs(resp) + return answer + + def parse_inputs(self, messages: List[dict]): + assert isinstance(messages, list), \ + 'Inputs must be a list of dictionaries with keys from ["system", "question", "answer"].' + prompt = '' + question = messages.pop[-1] + assert len(question) == 1 and 'question' in question.keys() + question = question['question'] + for m in messages: + for k, v in m.items(): + if k == 'system': + prompt += f'''[INST] <> {v} <> [/INST]\n''' + elif k == 'question': + prompt += f'''[INST] {v} [/INST]\n''' + elif k == 'answer': + prompt += f'''{v}\n''' + else: + raise KeyError(f'Invalid key of message: {k}') + prompt = ' ' + prompt + ' ' + f' [INST] {question} [/INST]' + return prompt + + def parse_outputs(self, response): + return response['choices'][0]['text'] + + @staticmethod + def supported_model_names(): + models = { + 'llama-2-7b-chat': { + 'hf_id': 'TheBloke/Llama-2-7B-GGML', + 'filename': 'llama-2-7b.ggmlv3.q4_0.bin' + }, + 'llama-2-13-b-chat': { + 'hf_id': 'TheBloke/Llama-2-13B-GGML', + 'filename': 'llama-2-13b-chat.ggmlv3.q4_0.bin' + } + } + return models + + @property + def shared_type(self): + return SharedType.Shareable diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000..de714be --- /dev/null +++ b/requirements.txt @@ -0,0 +1,2 @@ +llama-cpp-python +huggingface-hub